Reflecto

Single-layer X-ray reflectivity (XRR) analysis: a 1D CNN predicts film thickness, roughness and SLD from a measured curve (2θ, intensity .dat), then GenX refines a film-plus-SiO₂-on-Si model starting from that guess. The GenX script puts the SiO₂ layer above the film while the training data had it below; see #3. The background parameter is fitted but not applied; see #2.

The bundled weights (src/reflecto/assets/best.pt, stats.pt, config.json) come from experiment exp07 in ReflectoExp as of December 2025: config.json has the values of exp07's config.py at ReflectoExp b93c91f (2025-12-23: exp_name "cubic", fixed SiO₂ SLD 18.8, roughness 0-15 Å), not those of its current config.

Reflecto is one of the XRR research branches (CNN guess, then GenX differential evolution). ReflectoLearn carries an adapted copy of the GenX fit as reflectolearn.genx_fit (ReflectoLearn 6e160db). No comparison of the two approaches has been run, and the electron-density profile and density-ratio output here have no counterpart there.

Requirements

Python 3.13+, numpy, torch, genx3, bumps, pandas, matplotlib, tqdm (pyproject.toml, uv.lock).

Usage

uv sync
$env:PYTHONPATH = "src"     # pyproject 에 build-system 이 없어 패키지가 설치되지 않음
uv run main.py              # test_data/In2O3_Ar_30nm/#2_xrr.dat 분석 (저장소 루트에서 실행)

main.py 는 In₂O₃ 30 nm 시료로 AI 예측 → GenX 피팅 → 벌크 밀도 대비 비율 · FOM 출력 → 반사율 곡선과 전자밀도 프로파일을 그린다. 다른 파일은 main.py 의 data_dir / file_path 를 수정.

import reflecto
from reflecto.utils.math_utils import i0_normalize

df = reflecto.read_dat("sample.dat")  # 2열: tth, intensity (공백 구분, # 주석)
q = reflecto.tth2q(df["tth"].to_numpy())
R = i0_normalize(df["intensity"].to_numpy())
result = reflecto.analyze(
    q, R, fit=True
)  # result["prediction"], result["fit_params"], result["fom"]
reflecto.plot_analysis_result(q, R, result)

Structure

src/reflecto/
  core.py            analyze(): CNN 추론 + GenX 피팅
  engines/           CNN 모델 · 추론(inference.py) · GenX 피팅(fitting.py, script_builder.py)
  simulate/          GenX 반사율 시뮬레이션, 노이즈, SLD 프로파일
  utils/             .dat 읽기, 2θ→q, SLD ↔ 밀도 변환, GenX 층 상수
  assets/            학습된 가중치 (best.pt ≈ 64 MB)
test_data/In2O3_Ar_30nm/   측정 데이터, GenX .hgx, 결과 그림용 스크립트